Blog

Notes on building reliable AI products.

Atria Dawn Preview Turns Open Weights Into a Harness Test

Atria Dawn Preview opens GLM-5.2-based agent weights, but harnesses, verification, and runtime budgets will determine builder outcomes.

Apple’s Siri AI beta makes app actions, personal context, permissions, and PCC limits the new OS-level battleground for builders.

Mistral Vibe 2.25.4 fixes six shell-permission CVEs. The deeper lesson is how coding agents must bind approvals to actual runtime effects.

Anthropic’s adviser bundle shows how regulated AI will ship: workflow skills, uneven connector authority, human review, and evidence-led controls.

Anthropic’s embedded-evaluator pledge could shift AI assurance from model cards to operational audits. Here is what builders should demand next.

Reports link OpenAI agents to May’s RubyGems abuse. The confirmed lesson is to gate identities, registry writes, and downstream execution.

OpenAI priced GPT-Rosalind for October 5, but its internal-use API and preview Workbench make access, evidence, and governance the real product.

Sakana’s Fugu Max and Ultra v2 turn learned routing into an API product. Builders should measure accepted-task cost, drift, and opacity.

Anthropic’s September report shows why model identity, transcript custody, agent traces, and account revocation now belong in one security review.

Cognition’s SWE-2 turns Kimi K3 into a cost-aware Devin worker. Here’s how to read its benchmarks, effort levels and Terminal-Bench 4 gap.

OpenAI’s Agents API turns the Codex harness into managed infrastructure. Here is what it owns, what builders retain, and how to pilot it safely.

DeepSeek V4.1 Flash cuts input prices and will temporarily converge Pro routing. Here is what builders must re-test across caches, agents and APIs.